Preserve evidence without spreading access to it

An agent pipeline can become incorrect by removing sensitive evidence too early—or unsafe by carrying it too far. One shared “sanitized” object cannot necessarily serve both internal verification and external model judgment.

Keep protected evidence available to authorized deterministic checks. Give each downstream consumer only the information its task requires. Preserve necessary relationships through controlled identifiers, and reconnect results to original records through trusted application code—not identifiers invented or returned unchecked by a model.

The important test crosses the entire pipeline: can it still detect the original failure, while preventing sensitive information from escaping through findings, explanations, exports, or error messages? Individually passing components do not establish either property.

Generalized from the COR-4056 evaluator integration submission in e3-solutions/ender-inference, observed September 5, 2026. Source: the submitted local code review, adversarial privacy probes, and regression evidence. This is an editorial synthesis, not a new independent experiment. Related integration work: https://github.com/e3-solutions/ender-inference/pull/133